GLM-5.3 Marathon

GLM-5.3 Marathon is positioned for sustained, multi-step agent workloads rather than single-turn chat. Subconscious markets the model around an effective context of several million tokens, targeting sessions that run across hundreds of steps, which makes it well suited to long-running coding agents and autonomous workflows that would exhaust shorter-context models. The serving layer exposes the model through OpenAI- and Anthropic-compatible endpoints, allowing existing SDKs and agent frameworks such as Rig, OpenCode, Claude Code, Codex, Cursor, Copilot, and Pi to point at it with minimal configuration changes.

Subconscious describes itself as an AI lab that enhances language models with its TIMRUN inference runtime and a complementary post-trained TIM family of models, and the Marathon tier is the deployment profile in which GLM-5.3 is delivered for extended-context use. The model supports reasoning, tool calling, temperature control, structured output, and open-weight availability, so it can be deployed flexibly by teams that want to run, fine-tune, or self-host the weights while still benefiting from a managed long-context inference path. Practically, it fits builders who need an open base model exposed through standard APIs for agentic products that accumulate large working memories over long sessions.